This post originally appeared on IEEE Spectrum's Automaton blog.
At this point, seeing robot cars pop up at places like CES is getting less and less surprising and more and more frustrating as we think about just how many hurdles these vehicles have to drive over before we'll actually get to start using them. Don't get me wrong, there's been substantial progress, but when researchers say things like "hey we can make a car autonomous for $150"—as U.K. researchers said about their recently unveiled RobotCar project —it's time to get excited (and frustrated) all over again.
This car, a modified Nissan Leaf, comes from Oxford University. The first thing to know about it is that it doesn't use GPS. Really, this shouldn't come as a shock: GPS is good enough to tell you what road a car is on, but usually not whether a car is driving on the right (or left) side of that road, which is arguably more important. And unlike Google's self-driving car, the British vehicle doesn't rely on an expensive LIDAR sensor. Instead, Oxford's RobotCar relies entirely on scene recognition (from cameras and lasers), matching live imagery with a pre-existing database to figure out where it is, where to go, and what not to smash into. Here's how the car sees things:
With a system like this, you've got two different things going on at the same time. You've got a static information, which you get from maps that you've made of your route, and then you've got dynamic information, which includes other cars, traffic lights, bicyclists, pedestrians, Roombas, dinosaurs, wayward colonials, or whatever else you're likely to find on or near a British road. Here's how the researchers describe the difference:
Static information consists of semantic information like the location and type of road markings and traffic signs, traffic lights, lane information, where curbs are, etc. This kind of information rarely changes and so a fairly accurate model can be built before the vehicle actually goes out. And it will last. Of course, you don’t really want to blindly believe such a prior map for all time—after all, things do change when conducting roadworks, for example—but knowing where you can expect to find certain things in the world is already incredibly helpful. The prior semantic map will get updated over time with information the vehicle actually gathers out there in the real world.
Dynamic information relates to potential obstacles which are either moving or stationary: cars, bicycles, pedestrians, etc. Knowing where they are—and where they are likely going to be in the near future—with respect to the planned vehicle trajectory is crucial for safe operation as well as for appropriate trajectory planning. Dynamic obstacles are detected and tracked using an off-the-shelf laser scanner. The system scans an 85 degree field of view ahead of the car 13 times a second to detect obstacles up to 50 meters ahead.
Because the car relies on a static map, it's only able to operate autonomously in areas where that static map has been established. It can update the map on its own to account for minor changes, but it's not going to go out there and drive you someplace it's never been before all by itself. What it's really designed for is taking over on routes that you drive often: like, commuting to work. It'll identify routes where it has high confidence and ask if you'd like it to take over, and then if it ever starts to get confused, it'll request that you take over again. If you don't, it'll stop itself. Of course, ten years down the line, it could just download maps recorded by other cars, and drive absolutely anywhere.
As we've seen with projects from companies like Google and Toyota, there are multiple levels of autonomy. Off the top of my head (and it's late at night, so cut me some slack), I can think of the following:
- Full autonomy: This is what Google has: the car drives itself, and can react to changes and emergency situations. Driver optional.
- Restricted full autonomy: Stanford's Audi TTS is fully autonomous, but only in specific situations: it can handle all kinds of roads that it has maps for, but not variables like traffic. Driver optional, but only where applicable.
- Emergency full autonomy: Toyota is working on this; the car will aggressively take over from you if it detects an impending accident. Driver optional, but only in emergency situations.
- Highway assisted autonomy: Volvo has this system operational in Europe. Passenger cars autonomously follow a truck on a highway in a road train formation. Driver required to lead the train, and in cars for entry and exit.
- Highway driver assist: You can buy this system in luxury cars; it includes adaptive cruise control on highways and sometimes lane drift warnings. Driver required to be paying attention.
- Emergency driver assist: This system is also available in some luxury cars; if an imminent collision is detected, the car will autonomously apply brakes. Driver required the rest of the time. It's also worth mentioning that anti-lock brakes are a very primitive semi-autonomous emergency driver assist system.
- Emergency driver notification: I guess this one may not belong in the list, but here it is anyway: some luxury cars will track driver attention and fatigue and provide notifications if a dangerous situation arises.
The RobotCar project fits in somewhere between "Restricted full autonomy" and "Full autonomy," in that it can drive itself without needing you as long as it's familiar with the route. However, you legally still have to pay attention. This makes the car much, much safer, but it also means that you don't get to kick back and do something more productive, which is the fantasy we all have for autonomous cars.
Now, about this $150 price ... Well, here's the quote:
It could be only 15 years before self-driving systems become commonplace in cities as the price of installing the systems drops: "At present it costs about £5,000, but we're working to reduce that to £100," he said.
Along with the continual dropping price of hardware, it's worth remembering that you don't need expensive hardware if you can use clever software to make up for deficiencies in cheaper stuff. Getting from $7,700 down to $150 is more than an order of magnitude, though, and we shouldn't expect it to happen overnight. Luckily, or unluckily, it's going to take some orders of magnitude to make autonomous cars legal for daily consumer use anyway, so maybe by the time technology like this is commercially available, it'll also be cheap enough for us to use.